An Efficient Super-resolution Algorithm Based On Information Distillation

Image super-resolution technology uses the information of low-resolution images to reconstruct high-resolution images and is widely used in face recognition,target detection,and other fields.Single image super-resolution(SISR)combines the power of convolutional neural networks(CNN)to achieve the best super-resolution algorithm performance.However,as the network depth and width increase,CNN-based super-resolution methods face the problem of excessive inference latency in practical use.Exploiting feature redundancy in images is an essential feature of efficient CNN algorithms,and many lightweight SR algorithms also utilize feature redundancy to improve algorithm efficiency.To solve the above problems,this paper research the efficient superresolution algorithm based on information distillation.The main work and innovations are as follows:1.This paper analyzes the feature redundancy in super-resolution algorithms and proposes a fast and accurate SISR model by exploiting feature redundancy in spatial and channel dimensions,called spatial and channel feature distillation network(SCFDN).Specifically,SCFDN first splits the channels along the channel dimension and divides the feature map into flat features and detail features to remove redundant computations in the channel dimension.Then SCFDN performs feature distillation on the detail features to remove the redundant computation of the spatial dimension,then further processes the distilled detail features,and finally fuses the flat features and the detail features to reconstruct a high-resolution(HR)image.2.This paper also proposes to use the inverse sub-pixel convolutional(ISPC)layer as a downsampling layer to map the feature map from the spatial dimension to the channel dimension,thereby improving the effect of feature distillation.3.This paper proposes residual dense block(RDB)with high algorithm effect,high algorithm efficiency and low memory usage,which is used as the backbone network of the algorithm after forming stacked blocks.4.The experimental results confirm that compared with the current optimal SR algorithm,the SCFDN proposed in this paper is comparable in terms of algorithm effect and has better performance in terms of running time.